Aleph Alpha releases Kolibri, a German-focused model for public sector and industry

A mixture-of-experts model built for German

Kolibri combines 78 billion total parameters with around 3 billion active per token, according to Aleph Alpha. Its mixture-of-experts architecture selectively activates parts of the model during inference, which is how the company reaches the lower active parameter count.

German is central to the design. Kolibri is trained to reason in German, and German-language text makes up around 23% of its pre-training data. The model also uses a tokenizer tailored to German, designed to represent German text with fewer tokens and so support more efficient processing of German-language content.

Samuel Weinbach, Co-founder and Co-Chief Research Officer, said: "We designed Kolibri around the needs of German-language applications, from a tokenizer optimized for German to training the model to reason in German."

Agentic workflows, RAG and reasoning controls

Aleph Alpha says Kolibri supports agentic workflows, retrieval-augmented generation (RAG) and native tool calling. In RAG applications, the model uses supplied documents to answer questions. It is trained to abstain when those documents do not contain sufficient evidence.

Built-in reasoning controls let users adjust the balance between reasoning effort and response speed. The model is designed for deployment on infrastructure that customers control, with documented data provenance and design decisions.

Documented training-data governance

A large part of the announcement concerns compliance. Aleph Alpha says its technical report details measures addressing copyright, data protection and EU AI Act requirements, including training-data screening and checks on third-party datasets.

• All training data was screened against a blocklist of more than 4.5 million URLs, drawing on sources including the European Commission's Piracy Watch List.
• Each third-party dataset was assessed for license terms, lawful sourcing and adherence to opt-outs before use.

The model, its applicable license terms and supporting documentation are available on Hugging Face. Further details on architecture, evaluations and training methodology are in the Tech Report and a company blog post.

Sovereignty and the Cohere agreement

Ilhan Scheer, CEO of Aleph Alpha, framed the release in terms of sovereignty. He said: "For us, AI sovereignty means freedom of choice by retaining the ability to build and advance this technology, and giving customers control over how they use it." He added that the company's ambition is to bring that capability into everyday operations in industry and public administration.

Aleph Alpha also noted that it recently announced an agreement with Cohere to build a transatlantic sovereign AI company. The transaction remains subject to regulatory approval, and until closing Aleph Alpha continues to operate independently.

For organizations in German public administration and industry, the release offers a model aimed at German-language work that can run on infrastructure they control, with its training-data practices documented. The source does not give benchmark results or pricing, so how Kolibri compares with other models is left to the technical report and independent evaluation.